[GitHub] kurt-o-sys commented on issue #7349: time series delay output mxnet

[gi...@git.apache.org]

kurt-o-sys commented on issue #7349: time series delay output mxnet
URL: https://github.com/apache/incubator-mxnet/issues/7349#issuecomment-321056587
 
 
   Thx for you answer... It is a one-to-one RNN indeed, unfolded, I'd like to have a 'many-to-one'. Every time frame, a 'summary' of the events are added to the AI system. Here's how the network (more or less) looks like. 
   
   ![image](https://user-images.githubusercontent.com/2430465/29090285-50aea288-7c7f-11e7-8efa-126e545ac861.png)
   
   If I would unfold it, in the scenario described above, I'd have 51 times the same structure next to each. I would like to have only 1 output (the last one) during training. Setting to 0 is not really an option, since this would intervene with the training: a value of 0 is not the same as something that has more a meaning as 'not available yet'. 
   (During usage of the trained network, the errors made during the first outcomes are less important. It's just that at the start of the 'game', one just knows that the game state is still unsure. Only after a number of iterations, this improves towards how the game actually evolves).
   
   I still wonder how to achieve it, or stated another way, how I can 'remove the first x elements of the sequence_length' in a one-to-one RNN.
 
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


With regards,
Apache Git Services